by , ,
Abstract:
Domain randomization is a technique to bridge domain gaps by varying scenarios in a simulated environment to achieve generalization. It is an approach that can be implemented for reinforcement learning-based control applications in the aerospace domain, which is safety-critical and must be robust against external perturbations. The issue arises as machine learning deployment does not guarantee convergence, hence it is prone to error and instability. Furthermore, when the algorithm is trained in a simulated domain, it can be especially tricky to reach generalization when transferred to another domain with an underlying domain gap. This paper proposes a strategy to test the adaptability and reliability of reinforcement learning-based control systems on a realistic scale by conducting Sim2Sim transfer between Simulink textregistered, PyBullet, and MuJoCo, bridged with domain randomization. The experiment shows that domain randomization successfully reduces the simulation gap that exists between simulators for some states, backed with mean aggregation weights and bias ensemble generated from two different simulations from source domain to reach the best results.
Reference:
Phoebe Calista Lydwina, Takehisa Yairi, Naoya Takeishi:Domain Randomization for Sim2Sim in 6 DoF Rendezvous-Docking, In The 35th International Symposium on Space Technology and Science, Tokushima, 2025.
Bibtex Entry:
@conference{lydwinaISTS2025,
title = {Domain Randomization for Sim2Sim in 6 {DoF} Rendezvous-Docking},
author = {Phoebe Calista Lydwina and Takehisa Yairi and Naoya Takeishi},
labauthor = {Phoebe Calista Lydwina and Takehisa Yairi and Naoya Takeishi},
url = {https://ists.ne.jp/the35th/},
year = {2025},
abstract = {Domain randomization is a technique to bridge domain gaps by varying scenarios in a simulated environment to achieve generalization. It is an approach that can be implemented for reinforcement learning-based control applications in the aerospace domain, which is safety-critical and must be robust against external perturbations. The issue arises as machine learning deployment does not guarantee convergence, hence it is prone to error and instability. Furthermore, when the algorithm is trained in a simulated domain, it can be especially tricky to reach generalization when transferred to another domain with an underlying domain gap. This paper proposes a strategy to test the adaptability and reliability of reinforcement learning-based control systems on a realistic scale by conducting Sim2Sim transfer between Simulink textregistered, PyBullet, and MuJoCo, bridged with domain randomization. The experiment shows that domain randomization successfully reduces the simulation gap that exists between simulators for some states, backed with mean aggregation weights and bias ensemble generated from two different simulations from source domain to reach the best results.},
booktitle = {The 35th International Symposium on Space Technology and Science, Tokushima},
lang = {en}
}